Data centers are also responsible for the cloud and digital content. AI obviously motivated a recent explosion in data centers, but (possibly excluding non-AI parts of the overall apps) it’s not the majority of what data centers do, at least not yet. Grabbing a blurry screenshot from a McKinsey projection:
EDIT: Actually a more recent projection says otherwise
So what you’re saying is … I’m ruining the environment even responding to this post right now?? 
Thanks for your original post @mihalybaci ; I stand by you and fully share your concerns. There are also the ethical considerations of how these tools came to be, including dubious lacks of attribution and other shady practices that optimize performance. There are also ethical considerations regarding ongoing research regarding the impact of these tools on young developing minds. And also the corrosion of community aspects that is discussed on the sister thread to this one. We’re not discussing these aspects here (and that’s okay, this is for another post), but as you said it is really important to take the cosmic perspective from time to time. Not everything in this world should revolve around MORE-MORE-MORE: buy MORE, spend MORE, eat MORE, make MORE PRs, submit MORE prompts, use MORE tokens.
This blog post also corroborate the figure of heavy, yearly agentic use being roughly equal to one flight. From my estimate it’s also in the same ballpark as eating beef regularly. That’s not to say we shouldn’t have restricted all these things like 30 years ago, but I think it shows it’s not an issue that is specific to agentic workflows.
The world we live in is not as simple as many believe.
I’m not sure in what far e.g. China can be considered capitalistic. The developments there may be driven by private companies, but the strategy is defined by the communistic government.
Agreed. But if there was one thing which I’d like to pour the gasoline onto, it would be the pace of scientific research. If agentic coding could be used to accelerate our understanding of the world, our ability to confront and potentially mitigate or solve some of the enormous problems we’re facing as a species, wouldn’t that potentially be worth the cost, just as flying to JuliaCon could be worth the jet fuel?
It’s one of the reasons why the projects I’ve been working on are here, in Julia, not somewhere else.
I disagree, since we’ve been comparing the size of various impacts. I feel it’s quite relevant to point out that there are other optional activities which people engage in that are many orders of magnitude more damaging. Plus even the most annoying LLM output is better than being bombed. Under the assumption of a limited supply of outrage per capita, might as well place it where it counts?
I don’t see how that makes my comment false? We worked 14 hour days because we lacked productivity. Once productivity started to rise were able to buy leisure time without losing wages. It’s the industrial revolution that brought welfare, not labour laws. Although labour laws definitely helped establish rights for workers, once productivity increases made it possible to pass such laws. In many poorer countries people still have to work 12 hour days, because they lack productivity (which is obviously not the labourer’s fault)
Is that what it is? Or is it because they’re being exploited by a market system designed to do so?
There have been studies of indigenous foraging tribes in which labor required for an “affluent” life was 15-20 hours per week. That was without all this technology as well! Shouldn’t the technology increase the productivity and we should have more free time?
But who has been rewarded in the end? Didn’t Keynes write an essay on this, “Economic Possibilities for Our Grandchildren” (1930)? In which he forecasted that by 2030 productivity would grow enough to largely solve the economic problem, reducing the work week to about 15 hours and leaving humanity’s main challenge as what to do with its abundant leisure?
How’s that working out?
Are the shorter working hours (gosh, only 40 hours per week – if you’re lucky?!) something that just ‘emerged’ out of the primordial soup of the industrial era? No. They were fought for, and paid for in blood.
As an example, May 1, 1886, hundreds of thousands of US workers walked off the job in a coordinated national strike demanding shorter working hours. In Chicago it led to the Haymarket ‘affair’, strikes, marching, a labor demonstration which resulted in multiple deaths including strikers being shot by police. International Workers’ Day (May Day) is now recognized around the world, in part due to this event. (Haymarket affair - Wikipedia)
Yes low productivity is the core of the problem. It shows up everywhere: How Do Hours Worked Vary with Income? Cross-Country Evidence and Implications - American Economic Association
People are often self employed and have to work longer hours to compensate for low productivity.
I disagree, I think over-exploitation is the core, not productivity. The paper you cite shows this. It finds that in the richest countries, more highly paid workers work as much or more hours as wages increase, not fewer. If low productivity caused long hours, the most productive workers on earth would work the least. They don’t, unless they choose to, because they have that freedom.
What actually traps people in long hours is having little leverage. If you’re at the bottom of the socioeconomic ladder you’re the most vulnerable to exploitation and you can’t walk, so you take whatever terms you’re given. Plus the wage scale is pushed as low as possible, such that the only option is to work more to scrape by. The affluent don’t work less because they’re more productive, they work less because they can pick and choose, lean on their networks, grab the best-paid work and leave the scraps.
The freedom at the top and the grind at the bottom aren’t two separate outcomes, they’re the same system. That’s not a productivity gap, it’s a power gap. 
But I feel that this particular aspect has now fully veered away from the original topic, so I’ll leave it here for now.
From the paper:
We document that, on average, adults in the developing world work about 50 percent more hours per week than adults in rich countries. Average hours worked are higher in developing countries both for men and for women, for all age and education groups, and along both the extensive and intensive margins. Within countries, hours are decreasing with income on average
and particularly so in the poorest countries.
So regardless of the personal circumstances someone is in, on average in poorer countries they will still work more hours. The only explanation for that result to hold across all population groups is that productivity is higher and affording you to work less.
Because you have lower productivity. You can’t lose hours because every hour missed could mean no food on the table. And you have no leverage because if you lose your job there is no support network either, because you did not have the productivity to afford that. I would even go so far as to say that without substantial productivity increase, you would not be able to afford better worker protections because people would still be poor.
Going back to AI, the only reason I mentioned productivity is because we may see another wave of productivity increases similar to the industrial revolution, affording us to work even less hours. Which ultimately makes the contrast between rich and poor countries even greater than it already is. That is ofcourse only if the promise of AI holds up in the long term. And I’m not saying that I’m fully onboard the AI hype train and that I trust the stories sold to us by AI vendors trying to sell their greatest models. But I am basing this on the changes I am seeing around me, with more people picking up AI the effect becomes clearer. (I mostly just find it an interesting topic to think about, in my country people have already started working 4 days a week a while ago, even my dad who is a mechanic now works 4 days a week since 2022 without loss of any wages!)
The argument you’re presenting is dangerously close to wading into logical fallacy territory. Are data centers the number one source of CO2? No, and I never said they were. Are there other effective ways to reduce air pollution? Yes, and I never said there weren’t.
Is war bad? Yes.
But the argument of “look at all these other bad things! See that makes what I’m doing less bad!” does not hold water.
The whole point of my post is that people who take climate change and deforestation and wild lands seriously need to consider LLM usage in the same category as driving cars, eating less beef, and turning off all the lights when leaving the apartment/house.
And remember the number of data centers is increasing. The amount of power they consume is increasing. The number of users is increasing. If tech CEOs get their way, the number of users will be greatly increasing.
So, as I’ve been saying, the contribution by an individual to the negative environmental impact may be small, but the collective effect is large and getting larger, and the more important it becomes to discuss it.
My point is more about the scale of those things.
So if I drive half as many miles as average, or get a car that’s twice as efficient, and adopt that usage for a single year, then go back to normal, I should be able to use agentic coding heavily for about 10 years and be net neutral with just driving an average car an average amount that entire time. I drove only about 2000 miles last year, so that gives me at least a couple decades of AI use that I don’t need to feel guilty about.
Since I’ve never taken an international flight in my life, my carbon footprint is already lower compared to those who have and feel just fine about themselves, that I can also feel fine using these tools for the next few decades.
Low productivity, lack of institutions and/or low trust in institutions, low legal framework, poor access to natural resources, etc.
The specific part I claim is false was
Capitalism is a powerful force, but it is not an irresistible one.
In order to keep the Earth from warming further, less CO2 needs to be in the atmosphere. Carbon trading and swapping one source of CO2 for another isn’t going to get us there, only permanent reductions will.
Then use them if you feel fine with them. This isn’t about just you.
Perhaps in the sense of strict estimated CO2 emissions it may be somewhat analogous, but overall it is not a good analog at all.
Modern commercial air travel is a commodity with a wildly different ontology and embedding in our current economic, sociopolitical, and geopolitical realities.
LLM’s are being used, primarily by a small group of billionaires running AI/VC firms, to materially rearrange world economics, long-term power infrastructure and financing, water resource allocation, and their general purpose/sales pitch is to massively reduce labor costs in line with the material history of capital deployment of technology development/improvement. They are also widely being deployed in anti-democratic surveillance technology and in kill chains for illegal imperial wars and genocide (https://ai-killchain.airwars.org/).
To reduce the broad externalities of LLM’s to current noisy estimates of current CO2 emission intensity is, in my opinion, an extraordinarily reductive viewpoint.
I agree with @kahliburke to not really continue this little quagmire, though I think it is worth noting: this is not how logic or empiricism work. There is not only one possible explanation for this extremely general and simplified trend that you’re commenting on. Solving/explaining empirical problems such as this requires painful enumeration of all feasible material causes and investigating them. This is what @kahliburke has tried to allude to.
Um, my analogy here was specifically about climate impacts and environmental externalities.
I similarly don’t think garage door openers are a good analog overall, either.
Yes, you are right, my apologies. I have simplified my arguments because the discussion was getting ridiculous. I feel like I’m defending something which is so obviously true that I feel like I’m having a stroke reading these messages. I poured over chart after chart, ‘did I get it wrong?’, nope, they all show the exact same pattern. So I’m going to leave it at that.
